Power Considerations for Sensor Networks Mani Srivastava UCLA In - PowerPoint PPT Presentation

Power Considerations for Sensor Networks Mani Srivastava UCLA In collaboration with: USC/ISI (Bob Parker, Brian Schott) Rockwell Science Center (Charles Chien)

Outline � Power analysis of some sensor nodes � Energy-efficient multihop packet forwarding � Power-aware real-time processing

Power analysis of sensor nodes: Where does the power go? � High-end sensor node: Rockwell WINS nodes � StrongARM processor � Connexant’s RDSSS9M 900MHz DECT radio (128 kbps, ~ 100m) � Seismic sensor � Low-end sensor node: Experimental node similar to Berkeley’s COTS motes � Atmel AS90LS8535 microcontroller � RF Monolithic’s DR3000 radio (2.4, 19.2, 115 kbps, ~ 10-30m) � No sensors (but microcontroller has ADC)

Power Analysis of Rockwell’s WINS Nodes (Measurements) Processor Seismic Sensor Radio Power (mW) Active On Rx 751.6 Active On Idle 727.5 Active On Sleep 416.3 Summary Active On Removed 383.3 Sleep On Removed 64.0 � Processor Active Removed Removed 360.0 � Active = 360 mW Active On Tx (36.3 mW) 1080.5 � doing repeated transmit/receive Tx (27.5 mW) 1033.3 � Sleep = 41 mW Tx (19.1 mW) 986.0 � Off = 0.9 mW Tx (13.8 mW) 942.6 Tx (10.0 mW) 910.9 � Sensor = 23 mW Tx (3.47 mW) 815.5 � Processor : Tx = 1 : 2 Tx (2.51 mW) 807.5 � Processor : Rx = 1 : 1 Tx (1.78 mW) 799.5 Tx (1.32 mW) 791.5 � Total Tx : Rx = 4 : 3 at Tx (0.955 mW) 787.5 maximum range Tx (0.437 mW) 775.5 � comparable at lower Tx Tx (0.302 mW) 773.9 Tx (0.229 mW) 772.7 Tx (0.158 mW) 771.5 Tx (0.117 mW) 771.1

Power Analysis of Experimental Node (Measurements) Mode Power Level OOK @ 2.4 kbps OOK @ 19.2kbps ASK @ 2.4 kbps ASK @ 19.2kbps Tx 0.7368 14.88 15.67 16.85 17.76 Tx 0.5506 13.96 14.62 15.80 16.85 Tx 0.3972 12.76 13.56 14.75 15.54 Tx 0.3307 12.23 13.16 14.35 15.15 Tx 0.2396 11.43 12.23 13.43 14.35 Tx 0.0979 9.54 10.35 11.56 12.36 Rx 12.5 12.50 12.50 12.50 Idle 12.36 12.36 12.36 12.36 Sleep 0.016 0.016 0.016 0.016 2.4kpbs OOK Note 20 2.4kbps ASK Consumed Power (mW) 19.2kbps OOK � All powers in mW 18 19.2kbps ASK � Microcontroller (with ADC) Receive Mode 16 � Active = 8.7 mW 14 � Idle = 5.9 mW 12 � Off = 3 µ µ W µ µ 10 8 1 2 3 4 5 6 Tx Power Level

Some Observations from Power Analysis � In WINS node, radio consumes 33 mW in “sleep” vs. “removed” � Argues for module level power shutdown � Tx and Rx power � Rx power within 40% of maximum Tx power � Under certain circumstances, Tx power < Rx power! � Argues for: � MAC protocols that do not “listen” a lot � Low-power paging (wakeup) channel � Processor power fairly significant (30-50%) share of overall power � Sensor transducer power negligible � Use sensors to provide wakeup signal for processor and radio

Energy-efficient Multihop Packet Forwarding Architecture � Problem: radios often simply relay packets in multihop network � Traditional approach: main CPU woken up, packets sent to it across serial bus � power hungry computing and communication operations � Our approach: exploit programmable micro-controller in the Communication Subsystem to handle common cases of packet routing � can also do operations such as combining of packets with redundant information …zZ Z Multihop Multihop Packet Communication Communication Rest of the Node Packet Rest of the Node GPS Subsystem GPS Subsystem Radio Radio Micro Micro CPU CPU Sensor Sensor Modem Modem Controller Controller Our Approach Tradit ional Approach � Key challenge: how to do it so that every new routing protocol will not require a new radio firmware � Solution: application-defined all-layer packet forwarding

Application-defined All-layer Packet Forwarding Communication Packet Classifier Subsystem GPS Application-Defined Matching Rules Radio Micro & Actions Packet Modifier Modem Controller � Packet-classifier and packet-modifier driven by application defined matching rules and actions � Matching rules: and/or expressions using =, <, >, range operators on arbitrary packet fields (offset, length) � Actions: accept, forward, drop, field increment/decrement etc. � Rules and actions operate on arbitrary packet fields (any layer) � fields specified as (offset, length) � only simple, common cases handled at the radio � for complex cases packet sent to the main processor � Expressiveness: implemented the following as test cases � Node ID-based addressing and routing (IP-like) � Point-cast (send to a rectangular geographical area specified as destination) � Current proof-of-concept prototypes � Rockwell node � Experimental platform using Triscend’s microcontroller with on-chip FPGA

Power-aware Real-time Processing with Energy-fidelity Trade-off � Question: what kind of dynamic power management techniques makes sense on the CPU of the sensor node? � Software typically organized as RTOS with priority-based preemptive scheduling � Typically static priority � eCos, uCos, Neutrino, WinCE etc. � Traditional approaches of power management aren’t the best: � Select a (fixed) lower voltage when designing the board � Can’t handle tight deadlines � Shutdown whenever idle � Only gets a linear improvement � Example: consider a task set {(10, 3, 10), (14, 7, 14)} � CPU utilization is 80% � Shutdown will save 20% power � Can’t slow CPU by 20% (& reduce V) as deadlines no longer met � Can do better by combining static voltage scaling and shutdown (22.5% saving in this example) Problem: Traditional approaches use WCET (worst case execution time) and aim for no deadline misses

Reality #1: Significant Variation in Execution Times � WCET : BCET is typically >> 1, e.g.: Program Description BCET WCET WCET/BCET DES Data Encryption 73,912 672,298 9.1 DJPEG JPEG decompression 128x96 color 12,703,432 122,838,368 9.7 FDCT JPEG forward DCT 5,587 16,693 3 FFT 1024-point FFT 1,589,026 3,974,624 2.5 Matcnt Summation of 2 100x100 matrices 1,722,105 8,172,149 4.7 Piksrt Insertion sort of 10 elements 236 5,862 24.8 Sort Bubble sort of 500 elements 13,965 50,244,928 3598 Stats Sum, mean, var of 2 1000-size arrays 1,007,815 2,951,746 2.9 � But, execution time variations in sensor data are not random � temporal correlation in underlying physical signal � can attempt to predict!

Reality #2: Sensor Applications Tolerant to Deadline Misses � Computation deadline misses lead to data loss � Packet loss common in wireless links � e.g. a wireless link of 1E-4 BER means packet loss rate of 4% for small 50 byte packets � radio links in sensor networks often worse � Significant probability of error in sensor signals � noisy sensor channels � Applications designed to tolerate noisy/bad data by exploiting spatio-temporal redundancy � high transient losses acceptable if localized in time or space If the communication is noisy, and applications are loss tolerant, is it worthwhile to strive for perfect noise-free computing?

Exploiting Execution-time Variation and Tolerance to Deadlines � Our strategy: predict execution time of task instance and dynamically scale voltage even more aggressively so as to minimize shutdown � Execution time prediction � learn distribution of execution times (pdf) � Tasks with distinct modes can help the OS by providing hint after starting � E.g. MPEG decode can tell the OS after learning whether the frame is P, I, or F � But, some deadlines are missed! � Adaptive control loop to keep deadlines missed under control � Typical result: 1.5-3x higher power saving compared to best conventional schemes with dynamic voltage, with < 1% deadlines missed � Provides adaptive power-fidelity trade-off

Power-aware RTOS Scheduler Implementation � RTOS predicts the remaining runtime (at max CPU speed) of a task instance � calculated whenever the task instance enters the system, or is preempted � based on run-times of previous instances of the task, and the run-time consumed so far � e.g. weighted mean � e.g. a coarse-grained discrete probability distribution of actual run time of each task is calculated, and used to calculate E[remaining_runtime | runtime_so_far] � adaptively adjusts a multiplicative factor dependent on recent deadline misses � Voltage scheduling strategy � Statically calculate a maximal global slowdown factor for the task set � Dynamically calculate a task-specific slowdown factor on context switch to stretch the task to its remaining WCET � Results � Power savings of up to x2 (over best known DPM schemes) with deadline misses from < 1% to ~ 10% depending on task sets

Current Status � Simulation tool for RTOS power management evaluation � PARSEC discrete event simulation language � Two communicating entities: � Task Generator � generates task instances with run times according to a trace or a distribution � RTOS � sets CPU speed by setting voltage and frequency � implements runtime predictor � Variety of task sets from literature � Note: non-predictive scheme is obtained by setting predictor to always return WCET – run time so far. � Implementation in eCoS RTOS � Running on Intel StrongARM evaluation boardwith frequency variation (but no voltage variation)